Sedation modulates fronto-temporal predictive coding circuits and the double surprise acceleration effect

Two important theories in cognitive neuroscience are predictive coding and the global workspace theory. A key research task is to understand how these two theories relate to one another, and particularly, how the brain transitions from a predictive early state to the eventual engagement of a brain-scale state (the global workspace). To address this question, we present a source-localisation of EEG responses evoked by the local-global task – an experimental paradigm that engages a predictive hierarchy, which encompasses the global workspace. The results of our source reconstruction suggest three-phases of processing. The first phase involves the sensory (here auditory) regions of the superior temporal lobe and predicts sensory regularities over a short timeframe (as per the local effect). The third phase is brain-scale, involving inferior frontal, as well as inferior and superior parietal regions; consistent with a global neuronal workspace (as per the global effect). Crucially, our analysis suggests that there is an intermediate (second) phase, involving modulatory interactions between inferior frontal and superior temporal regions. Furthermore, sedation with propofol reduces modulatory interactions in the second phase. This selective effect is consistent with a predictive coding explanation of sedation, with propofol acting on descending predictions of the precision of prediction errors; thereby constraining access to the global neuronal workspace.


Introduction
1 The term accelerate is primarily used metaphorically, although, the advanced P3 component that we identified in [Shirazibeheshti2018] (see figure 5 LxG panels in that paper) does in fact exhibit an acceleration in the mathematical sense. That is, the tangent to the curve (the velocity) changes rapidly with time, meaning that the rate of the rate of change with time (i.e. the acceleration) is bigger in the advanced P3 (the LDGD condition).
To answer these questions, we report a Multiple Sparse Priors [Friston2008] source localisation of the local-global task. This enabled us to characterise the neurophysiological trajectory of neuronal responses as they propagate from rapidly changing early responses restricted to sensory areas (for us in auditory cortices) to slowly changing late (c.f., metastable) responses (involving temporal, frontal and parietal areas). Our key finding was that the transition between these two phases -sensory-bound to global -involves a transient engagement of the superior temporal-inferior frontal network. Furthermore, the interaction between local and global violations on responses in this network is attenuated by propofol sedation.

Methods Participants
Originally, 22 neurologically healthy adults were included in the study, but two recordings were lost due to technical issues, leaving 20 participants (9 male; 11 female) (mean age = 30.85; SD = 10.98).

Experimental Design
The local-global auditory oddball task, devised by Bekinschtein [Bekinschtein2009], was used to characterise differences between local and global effects after healthy sedation and subsequent recovery. As shown in Figure 1, local regularity was established using sequences of five tones, or quintuples, where the last tone may or may not vary from the preceding four tones (local deviant versus local standard respectively). Global regularity was established as the most frequently presented quintuple type within a block, either local standard (all five tones the same) or local deviant (different last tone). Thus, violations in global regularity were expressed by the presentation of a quintuple that differed from the frequently presented type in any block. To ensure global regularity was established, a habituation period of 20 to 30 quintuples was presented at the beginning of the block. After the habituation phase, the ratio between the standard and deviant quintuples was set to 80/20. This created four conditions: (1) local standard / global standard (LSGS), (2) local deviant / global standard (LDGS), (3) local standard / global deviant (LSGD) and (4) local deviant / global deviant (LDGD) (see b, c, d and a in Figure 1). Quintuples comprised 5 tones of 50ms duration each, presented via headphones, with an intensity of 70dB and an SOA of 150ms. All tones were synthesised with 7ms rise and 7ms fall times. Participants were asked to count the number of global deviants they heard during both sedation and recovery phases of the study as an incidental task to reduce fluctuations in attentional set.

Sedation
During surgery or procedures for diagnosing medical conditions, it is common to take the patient to a sedative plane (as opposed to general anesthesia); also for pain control it is common to use propofol to relax the patient or take him/her/they to the point of sleep. In sedation research studies, the state is defined either by the target concentration in blood (light, medium or moderate sedation; light, medium or deep anesthesia) and /or a clinical responsiveness scale like the Ramsay 2 . However, in the study analysed in this paper, the sedation states are defined in an even more detailed manner, by exact concentration of propofol in blood in each state and by RTs and response misses in a behavioral task, as described for this experiment in [Chennu2016b].
In the experiment we are analysing, we knew that the target concentrations induced light to moderate sedation, where participants range from mild changes in relaxation to mostly unresponsive, but easily "arousable", since these dose-responses had been defined already, as in [Stamatakis2010, Adapa2014, Barttfeld2015, David2007].
Specifically, we knew the behavioral and drug in blood pattern, enabling us to place participants around "the verge of unconsciousness". The specifics of how we manage to maintain participants in this state, with such small doses are elaborated in [Absalom2009]. The local-global task was presented on two occasions; once during either mild (half of participants) or moderate (the other half) sedation and once 20 minutes later, when participants were considered to be in recovery (i.e. no longer sedated). Sedation in this study induced a heavily relaxed but behaviourally responsive state. All participants were tested both under sedation and subsequently in recovery, creating a repeated measures design. Each experimental run began with an awake baseline period lasting 25-30 minutes followed by a target-controlled infusion of propofol [Marsh1991], administered via a computerized syringe driver (Alaris Asena PK, Carefusion, Berkshire, UK). Three blood plasma levels were taken-0.6 μg/ml (mild sedation), 1.2 μg/ml (moderate sedation), and after recovery from sedation. A period of 10 minutes was allowed for equilibration of calculated and actual plasma propofol concentrations before cognitive tests commenced. Following cessation of infusion, plasma propofol concentration exponentially declined toward zero and approached zero in 15 minutes leading up to behavioural recovery. In light of this, the recovery condition started 20 minutes after cessation of sedation.
All procedures were conducted in accordance with the Declaration of Helsinki. The participants provided written informed consent and were healthy controls. Ethical approval for testing healthy controls was obtained from the Cambridgeshire 2 Regional Ethics Committee.

EEG recording
During pre-processing, two patients were excluded due to artefacts, therefore 18 participants were taken forward for analysis. Participants were asked to close their eyes during data collection to avoid eye artefacts in the data. EEG data were collected on two occasions: during sedation and then recovery. A Net Amps 300 amplifier (Electrical Geodesic Inc., Oregon, USA) with a high-density cap of 129 channels was used for data acquisition; and preprocessed data were obtained using custom Matlab scripts based on EEGLab. EEG timeseries were recorded in microvolts (μV), with a sampling frequency of 250Hz, and referenced to vertex (channel Cz). After recording, the data were segmented from -200ms Co-registration between the scalp level and the source level in MRI space was applied.
Coordinates at the scalp level were based on a standard GSN-Hydrocel template with 128 channels. Three fiducials were used to map the coordinates from sensor space to source space: nasion, left peri-auricular point and right peri-auricular point. A template head model based on the MNI brain was applied with a cortical template mesh of 8196 dipoles, which contains the coordinates of the dipole sources.
As previously stated, we discarded 36 electrodes on the face, neck and cheek from the montage, since they were noisy and dominated by muscle artefacts. This left us with incomplete coverage, making the localisation problem more difficult. As a result, we constrained the cortical mesh by selecting cortical areas that are, a priori, most likely to generate evoked responses; i.e., only regions of the temporal, frontal and parietal lobes were included for source reconstruction. Figs. 9 and 10 and supporting text in the supplementary material provide a detailed justification for this choice. We did not include deep sources or sources in the occipital and motor cortices, as there is no prior evidence suggesting these regions are related to the effects of interest in the present study.
The forward model was computed using the BEM (Boundary Element Method) [Phillips2007] as standard in SPM12 for EEG-based source reconstruction, with a three layer head model, i.e. skin, skull, and brain. At the source level, a mesh based on an MRI template is used to simulate the 1484 dipolar amplitudes in the brain.
We selected a time-frequency window for the source reconstruction. The frequency band used was the same as used in pre-processing with a range of [0.5 20] Hz. The window used for the source reconstruction is from 400ms after the first stimulus onset in a quintuple, to the end of the analysed segment. This included the baseline and the evoked response associated with the fifth tone up to the end of the quintuple.

Window placement for image extraction
After source inversion, two analyses were performed in order to characterise the spatial and temporal responses. The first used statistical parametric mapping to test for differences in evoked responses in all sources. The second analysis focused on responses in Regions Of Interest (ROI) based on regionally specific time-series at the source level. The ROIs were defined based upon effects identified by the statistical parametric mapping, as explained below.
Statistical parametric mapping summarizes the activity on the mesh, across the time window chosen. The evoked power was averaged within time windows for the frequency range [0.5 20] Hz. A spatial filter (FWHM=1mm) was used to smooth the dipole activity in three dimensional source space.
Evoked power under each condition was calculated as the root mean squared response (in source space) over the window. Statistical inference in this context, then, required the placement of time windows. Tailoring such windows post-hoc to the EEG data, risks biased sampling and could inflate false positive rates (e.g. [Brooks2016]

Statistical analysis Hierarchical spatial responses
For each participant (18) and each condition (8), three images of evoked power (one for each time window) were created for General Linear Model (GLM) statistical analysis. The experimental design can be summarized as a 2x2x2 within-subjects design, with 3 factors: sedation, local, and global. Each factor comprises 2 levels: sedation and recovery (for sedation); local standard (LS) and local deviant (LD) (for local); and global standard (GS) and global deviant (GD) (for global). The ensuing statistical analysis can be summarized as follows.
The first issue was to understand the relationship between local and global manipulations. To do so, we first looked at the local effect and the global effect individually and then the local x global interaction. Since the local and global effects have been extensively explored and are well documented in the literature; our analyses of these effects serve as sanity checks of our source localisation. That is, if the MSP algorithm localises these effects to the expected brain areas, we can have confidence that the reconstruction scheme can localise the effects for which there are fewer precedents.
The second issue was to understand the effect of sedation. We therefore assessed the main effect of sedation, the sedation x local interaction, the sedation x global interaction and the three way (sedation x local x global) interaction. We

Temporal dimension
Source analysis reveals which cortical sources exhibit significant effects of interest, within the different time windows. For illustrative (but not statistical) purposes, we investigated how activity changes through time at the source level, as follows. A source in each brain regionselected to plot the source time-series -was taken from the peak of the significant cluster, during the middle window [250,350] ms. Specifically, we selected the temporal lobe sources located at the peak of the cluster for the local effect, the frontal lobe sources located at the peak of the cluster in the local x global interaction, and the parietal sources located at the peak of the cluster for the global effect. arrow). The local effect is again significant in the temporal region during the middle window (solid blue arrow), whereas it is not significant in the late window. Fig. 2B) shows that frontal clusters are also significant in the middle window. This is shown by the time course of the frontal cluster in Fig. 2D), with an effect peak during the middle window. Table 1 summarizes the statistical results for each cluster in the early and middle windows. For each cluster, the peak location is described in the second column and the F-statistic for the peak of the cluster in the third column. The family-wise error (FWE) corrected p value is presented (4th columns) with the size of the cluster (5th columns). This shows a strong effect (P FWE < 0.001) for all the clusters shown in the table.

Global effect
The global effect is presented in Fig 3. In the early window, the global effect is significant in both left and right frontal sources, as shown in Fig 3C). This early global effect can be related to the contingent negative variation (CNV) [Chennu2013], which is usually observed at frontal electrodes. Accordingly, the time course for the frontal area in Fig 3F) shows where it also appears on the scalp.  Table 1. Statistics for each cluster for the early and middle windows. Each cluster, named in the first column, is characterized by its peak location in MNI coordinates as shown in the second column, the F-value for that peak (third column), the p-value (fourth column) and the cluster size (last column). The p-value highlights the significant cluster after family-wise error correction, set to an alpha of 0.05.
During the middle window, as shown in Fig 3D), the global effect is significant in the temporal, parietal and frontal regions with the time course represented respectively in Fig 3A), Fig 3B) and Fig 3F). Consistent with Wacongne et al 2011 and Chennu et al 2013, this window involves a broad network of brain activity indicated by the significant clusters. Fig 3A) shows the time course in the temporal area. The global deviant in this area diverges from global standard, but is significant only during the middle window, before disappearing in the late window. In the late window, Fig 3E), the global effect is significant in a network comprising both frontal and parietal regions. The time course for the parietal cluster is plotted in Fig 3B), which shows the global effect is significant during the middle and late windows with a peak at the beginning of the late window. Additionally, specifically on the left side, a second more dorsal parietal cluster appears which was not present in the middle window. Finally, Fig 3F) shows the frontal time course of the global effect, which is significant in all three windows, with a peak in the middle of the late window. This frontal area is the most activated for the global effect, starting from the CNV, until achieving the strongest effect during the late window.
The statistical results from SPM for each significant cluster are shown on Table 2

Local-by-global interaction
The local-by-global interaction is significant only in the middle window, as shown in Fig. 4B).
The left temporal time course is shown in Fig. 4A), with a small non-significant increase in the interaction effect (green line), which peaks after the early window. This is followed by a significant (P FWE = 0.003) second increase that is in the middle window. Clusters are also observed in the frontal area. Fig. 4C) shows the time course for the left frontal cluster. The interaction is significant in the middle window (P FWE = 0.009), with a positive interaction before a reversal of the effect (green line) in the late window, which does not reach significance. The details of the statistical results from SPM are presented in Table 3. This interaction in the middle window suggests that a fronto-temporal network is responsible for linking the local and global effects, and that these two effects (local and global) are not (strictly speaking) independent.

Three way interaction
Finally, the time-series for all conditions and the three-way interaction (local-by-global-bysedation) with its standard error are shown in Fig 5A). The three way interaction is significant in the late window, with its corresponding significant clusters in the frontal lobe shown in Fig 5D). To characterise the causes of the three way interaction, we explored the two simple effects interactions that constitute it. Specifically, the local-by-global interaction is presented separately for sedation and recovery in Fig 5B) and Fig 5C) respectively. Notably, the localby-global interaction was significant in the late window when participants had recovered, but not when they were sedated. Indeed, the local-by-global effect (green line) had opposite polarities when sedated and recovered for much of the late window. This difference between sedated and recovered seems to be carried by two properties. Firstly, the LDGD condition terminates more sharply when recovered, and secondly, the LSGD condition has a dramatically higher amplitude when recovered. The former is exactly consistent with the deceleration of the accelerated prediction error reported in [Shirazi2018], suggesting that inferior frontal regions are the source of this shifting neural responsiveness.

Figure 5. Three way interaction: A) source time-series for the frontal left cluster of the threeway interaction and the eight conditions involved. B) Local-by-global interaction source timeseries for the sedation conditions in frontal cluster. C) Local-by-global interaction source timeseries for the recovery conditions in frontal cluster. D) 3D glass brain of the significant clusters
for the three way interaction in the late window. E) 3D glass brain of the significant cluster for the local-by-global interaction when recovered.
The latter of these properties (amplitude increase for LSGD) is particularly striking, and important, since the LSGD condition is -in a sense -the most cognitively demanding  Additionally, we found a significant effect of sedation in the temporal region, and a significant sedation-by-local interaction in the frontal region, which are presented in the supplementary material (section "Further Sedation Effects"). The sedation-by-global interaction was not found to be significant.

Discussion
We have presented a source localisation of the neuronal responses evoked by the localglobal paradigm -and the effect of sedation with propofol on these responses. In this way, we have addressed how two key cognitive neuroscience theories -predictive coding and the global workspace -are related, and the neurophysiological correlates of this interrelationship.
Importantly, the two theories (predictive coding and Global Neuronal Workspace (GNW)) do already include some related concepts. In particular, prediction has been discussed within the GNW context, for example, the original formulation of GNW [Deheane1998] did discuss "anticipation" and pre-representation. Additionally, and perhaps most notably, [Wacongne2012] presents an important neural instantiation of predictive processing, which simulates the mismatch-negativity. However, we would argue that predictive coding goes further, by providing a full multi-level architecture of brain processing that is theoretically grounded in the mathematics of (generative) Bayesian inference [Friston2010], with concepts such as confidence in (i.e. the precision of) the prediction error to the fore. In this regard, predictive coding provides a full instantiation of the core hypothesis that the cortical architecture is configured to minimize the difference between bottom-up representations, driven from sensory input, and top-down representations of expectations, with this interpretation obtaining at all hierarchical levels. In this sense, there is a clear need to reconcile predictive coding and the Global Neuronal Workspace.
In respect of neural correlates, it is important to acknowledge the constraints associated with our source localisation analysis. In particular, we explicitly placed masks (which were justified by prior precedent; c.f. supplementary material section 'Subspace selection and mask placement') to constrain the source reconstruction. In this respect, it is not surprising that we found sources in these a priori regions. However, exactly where those sources fell, especially in the large regions of the frontal and parietal masks, is of interest. Of greater note, is how the MSP algorithm un-mixed variability amongst the regions and how that un-mixing progresses through the time-course of the evoked response. In this respect, a sanity check of our findings is that, as one would expect, early effects are temporal, with a following propagation out from this sensory region to frontal and parietal regions.
The local and global main effects we observed also have considerable face validity. In particular, as shown in Fig 2A,  1. three phases of processing (early, transitional and late); 2. fronto-temporal interaction between local and global in the transitional phase; 3. a failure to detect a main effect of sedation in parietal sources; and

a three way interaction between propofol and local and global effects
We elaborate on these in turn.
1) Three phases: Fig 6 depicts the neurophysiological realisations of the putative three phases. As discussed previously, the local effect manifests in the early window (early phase) in source space, very much as one would expect -expressed predominantly in superior temporal regions, which include auditory cortices. Additionally, a stereotypical global workspace is present in the late window (late phase). Importantly, our middle window (transitional phase) appears to exhibit qualitatively distinct effects, in terms of the set of sources involved and condition-specific effects exhibited. In particular, the local and global effects only interact in the middle window, suggesting a modulatory exchange between temporal and inferior frontal regions, where the local-by-global interaction was expressed. 2) Local times global Interaction and the transition phase: to elaborate further, the interaction between local and global in the middle window suggests a multiplicative or difficult to interpret 3 , and there remains the possibility that a more highly powered experiment would find an effect at parietal lobe.
Inference though can be less equivocal with regard to the significant effects involving the sedation factor. In particular, we observed a temporal source for the main effect of sedation, see Figs 7A), B) of the supplementary material, which, in the early window, is consistent with the known enhanced N1 in anaesthesia [Ypparila,2004], but for us was also seen in the middle window. Additionally, temporal and frontal sources evidenced a sedation x local interaction; see Fig 7E), F) of supplementary material. This provides some evidence that sedation modulates responses early in the processing pathway. The most interesting effect though was the local by global by sedation effect, which was significant in the late window. Secondly, the most striking feature driving the three way interaction at frontal in the late window is the dramatically higher LSGD condition when recovered than when sedated; see Figs 5B), C). Importantly, the LSGD condition is most dependent upon long-term temporal integration. In particular, this global deviance is not marked by a sensory prediction area (since it arises during a local standard quintuple). Thus, the deviance is not initiated by a strong bottom-up signal (i.e. a local prediction error). That is, it is a pure global deviance condition, with its detection intrinsic to higher hierarchical levels.

4) Three-way Interaction
It seems then that sedation impairs this capacity to detect deviance intrinsically at higher levels, at least at inferior frontal sources. This finding is in many respects consistent with the intent of the local-global task; i.e., to differentiate processing that requires temporally sustained integration over an extended period of time, and the role of consciousness in this temporally extended evidence accumulation. Thus, our findings provide suggestive evidence that, in respect of the action of propofol, reduced awareness diminishes long duration processing of temporal integration, supported by inferior frontal sources. In terms of predictive coding, this finding is consistent with a reduction in the precision of ascending prediction errors. This follows because precision corresponds to the rate of evidence accumulation (Hesselmann, Sadaghiani et al. 2010, FitzGerald, Schwartenbeck et al. 2015. In other words, belief updating in response to precise prediction errors converges more quickly than in the setting of imprecise prediction errors -or a pharmacological reduction in gain of responsiveness of populations encoding prediction errors. This particular effect of propofol would endorse the hypothesis we highlighted earlier that ignition depends upon a phase transition that itself rests upon attentional selection of ascending prediction errors that is mediated by the modulatory effects of predicted precision.

Sedation
The "sedation state" that we are exploring involves placing the participant at the fringe of consciousness, which may correspond to a (weakened, but) active bottom-up stimulus strength and lower top-down attention. In this regard, it may be comparable to the "preconscious" state in [Dehaene2006]. However, it is notable that in our data, sedation reduces, although seems not to eliminate, the sensory response, e.g. see supplementary material section "Further Sedation Effects", where Figure 7A,B,C shows a clear sedation effect at sensory areas. Because Event Related Potentials are averages across many trials, it is always difficult to know whether the reduction of a component is due to a consistent reduction at the single-trial level, or increased variability in the response. Thus, it is possible that the reduction in sensory components with sedation that we observe are caused by intermittent activation, in which sensory input is extremely weak, even absent, on some trials, and strong, In summary, although there remains considerable uncertainty, the key ERP components we observe are (increased) responses to unexpected/ deviant stimuli, which would naturally be considered a signalling of prediction error. In this context, we can give a more specific candidate interpretation of the effects of sedation: predictive coding suggests that the feedforward error signal reflects a prediction error weighted by its precision or confidence afforded that error signal. Confidence in our context would be driven by an assessment of the level of irreducible sensory prediction error -or by descending predictions of precision based upon belief updating higher in the hierarchy (see Figure 6). The two key effects of sedation that we observe are (1) a reduction in amplitude and (2)

Top-down or bottom-up ignition?
Our source localisation implicates a superior-temporal -inferior-frontal circuit in this modulation of response by sedation. Although -on the basis of the findings presented herewe cannot be certain of the direction of modulatory influence (descending or ascending) in this circuit. This is reflected in the two versions of the three phase theory presented in Fig 6. [Shirazi2018] proposed that the acceleration of the global response -due to coincidental local deviance -is caused by a feed-forward modulation from the local effect circuit onto the global effect circuit. This is the direction of modulatory influence presented in Fig 6, version 1, see panel V.1B. This might be considered an explanation of the acceleration by double surprise that sits most easily with the simplest line of temporal causation, since registration of local deviance would naturally be considered to precede registration of global deviance.
However, the other direction cannot be excluded. That is, it could be that a weakening of the modulatory influence of inferior-frontal on superior-temporal areas is what drives the slowed and attenuated responses observed when sedated. Such an explanation would be consistent with the second version of our three phase theory presented in Fig 6, where modulation is mediated in a feedback direction (i.e., descending predictions of precision), see panel V2.B.
From a predictive coding perspective, as previously suggested, a potential explanation of the effect of sedation is that it reduces the precision of sensory prediction errors, potentially carried by a feedback link from inferior frontal to temporal regions, as per Fig 6 (V2.B). This would effectively attenuate the gain on the ascending transmission of prediction errors. As noted in the introduction, this formulation of the transition from local to global processing (i.e., ignition of the global workspace) provides a graceful synthesis of predictive coding and global workspace theory that is grounded in neurophysiology (via gain control and neuromodulation) -while at the same time speaks to psychological concomitants of conscious processing (via attentional selection that accompanies perceptual synthesis over extended periods of time).